{
  "version": "v3",
  "generatedAt": "2026-05-19T17:58:04.815Z",
  "datasets": [
    {
      "id": "v3_Semi_Private",
      "displayName": "ARC-AGI-3",
      "url": ""
    }
  ],
  "evaluations": [
    {
      "datasetId": "v3_Semi_Private",
      "datasetDisplayName": "ARC-AGI-3",
      "modelId": "anthropic-opus-4-6-max-effort",
      "modelDisplayName": "Anthropic Opus 4.6 (Max)",
      "modelType": "CoT",
      "modelGroup": "anthropic-opus-4-6-max-effort",
      "modelReleaseDate": "2025-12-17T00:00:00.000Z",
      "providerId": "Anthropic",
      "providerDisplayName": "Anthropic",
      "providerColor": "#f93c32ff",
      "score": 0.0051,
      "cost": 8866.2,
      "resultsUrl": "",
      "display": true,
      "labelOffsetX": -100,
      "labelOffsetY": -15
    },
    {
      "datasetId": "v3_Semi_Private",
      "datasetDisplayName": "ARC-AGI-3",
      "modelId": "xai-grok-4-20-beta-0309-reasoning",
      "modelDisplayName": "Grok 4.20 (Beta Reasoning)",
      "modelType": "CoT",
      "modelGroup": "xai-grok-4-20-beta-0309-reasoning",
      "modelReleaseDate": "2026-03-05T00:00:00.000Z",
      "providerId": "xAI",
      "providerDisplayName": "xAI",
      "providerColor": "#ff1ed9",
      "score": 0.0009,
      "cost": 3775,
      "resultsUrl": "",
      "display": true,
      "labelOffsetX": -100,
      "labelOffsetY": -15
    },
    {
      "datasetId": "v3_Semi_Private",
      "datasetDisplayName": "ARC-AGI-3",
      "modelId": "openai-gpt-5-4-2026-03-05-high",
      "modelDisplayName": "GPT-5.4 (High)",
      "modelType": "CoT",
      "modelGroup": "openai-gpt-5-4-2026-03-05-thinking",
      "modelReleaseDate": "2026-03-05T00:00:00.000Z",
      "providerId": "OpenAI",
      "providerDisplayName": "OpenAI",
      "providerColor": "#1e93ffff",
      "score": 0.0021,
      "cost": 5187.41,
      "resultsUrl": "",
      "display": true,
      "labelOffsetX": -100,
      "labelOffsetY": -15
    },
    {
      "datasetId": "v3_Semi_Private",
      "datasetDisplayName": "ARC-AGI-3",
      "modelId": "google-gemini-3-1-pro-preview",
      "modelDisplayName": "Gemini 3.1 Pro (Preview)",
      "modelType": "CoT",
      "modelGroup": "google-gemini-3-1-pro-preview",
      "modelReleaseDate": "2026-03-05T00:00:00.000Z",
      "providerId": "Google",
      "providerDisplayName": "Google",
      "providerColor": "#4ecc30ff",
      "score": 0.0042,
      "cost": 2213.53,
      "resultsUrl": "",
      "display": true,
      "labelOffsetX": -100,
      "labelOffsetY": -15
    },
    {
      "datasetId": "v3_Semi_Private",
      "datasetDisplayName": "ARC-AGI-3",
      "modelId": "openai-gpt-5-5-2026-04-23-high",
      "modelDisplayName": "GPT-5.5 (High)",
      "modelType": "CoT",
      "modelGroup": "openai-gpt-5-5-2026-04-23-thinking",
      "modelReleaseDate": "2026-04-23T00:00:00.000Z",
      "providerId": "OpenAI",
      "providerDisplayName": "OpenAI",
      "providerColor": "#1e93ffff",
      "score": 0.0043,
      "cost": 10000,
      "resultsUrl": "",
      "display": true,
      "labelOffsetX": -100,
      "labelOffsetY": -35
    },
    {
      "datasetId": "v3_Semi_Private",
      "datasetDisplayName": "ARC-AGI-3",
      "modelId": "anthropic-opus-4-7-high",
      "modelDisplayName": "Opus 4.7 (High)",
      "modelType": "CoT",
      "modelGroup": "anthropic-opus-4-7",
      "modelReleaseDate": "2026-04-16T00:00:00.000Z",
      "providerId": "Anthropic",
      "providerDisplayName": "Anthropic",
      "providerColor": "#f93c32ff",
      "score": 0.0018,
      "cost": 10000,
      "resultsUrl": "",
      "display": true,
      "labelOffsetX": -100,
      "labelOffsetY": -15
    }
  ]
}
